On 02/03/2023 11:00, Andrzej Hajda wrote:
On 02.03.2023 11:43, Tvrtko Ursulin wrote:On 08/02/2023 10:51, Andrzej Hajda wrote:Write-combining memory allows speculative reads by CPU. ggtt->error_capture is WC mapped to CPU, so CPU/MMU can try to prefetch memory beyond the error_capture, ie it tries to read memory pointed by next PTE in GGTT. If this PTE points to invalid address DMAR errors will occur. This behaviour was observed on ADL, RPL, DG2 platforms. To avoid it, guard scratch page should be added after error_capture. The patch fixes the most annoying issue with error capture but since WC reads are used also in other places there is a risk similar problem can affect them as well. Signed-off-by: Andrzej Hajda <andrzej.hajda@xxxxxxxxx> Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>Research tells the explanation is plausible so: Acked-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx>Thanks for looking at it.On patch details below.What about "simiar risk in other places" - are there any plans to asses the potential impact elsewhere?Yes, merging this patch is the 1st step, as error_capture is responsible for most (maybe even all) regular DMAR read errors. After removing this noise it will be much easier to spot other DMAR read errors. There are also multiple regular DMAR write errors (less frequent, but still), which I hope to debug if time permits.Regards Andrzej--- This patch tries to diminish plague of DMAR read errors present in CI for ADL*, RPL*, DG2 platforms, see for example [1] (grep DMAR). CI is usually tolerant for these errors, so the scale of the problem is not really visible. To show it I have counted lines containing DMAR read errors in dmesgsproduced by CI for all three versions of the patch, but in contrast to v2I have grepped only for lines containing "PTE Read access". Below stats for kernel w/o patch vs patched one. v1: 210 vs 0 v2: 201 vs 0 v3: 214 vs 0 Apparently the patch fixes all common PTE read errors.In previous version there were different numbers due to less exact grepping, "grep DMAR" catched write errors and "DMAR: DRHD: handling fault status reg"lines, anyway the actual number of errors is much bigger - DMAR errors are rate-limited.[1]: http://gfx-ci.igk.intel.com/tree/drm-tip/CI_DRM_12678/bat-adln-1/dmesg0.txtChangelog: v2: - modified commit message (I hope the diagnosis is correct),- added bug checks to ensure scratch is initialized on gen3 platforms. CI produces strange stacktrace for it suggesting scratch[0] is NULL,to be removed after resolving the issue with gen3 platforms. v3: - removed bug checks, replaced with gen check. v4: - change code for scratch page insertion to support all platforms, - add info in commit message there could be more similar issues Regards Andrzej --- drivers/gpu/drm/i915/gt/intel_ggtt.c | 31 ++++++++++++++++++++++++---- 1 file changed, 27 insertions(+), 4 deletions(-)diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.cindex 842e69c7b21e49..6566d2066f1f8b 100644 --- a/drivers/gpu/drm/i915/gt/intel_ggtt.c +++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c@@ -503,6 +503,21 @@ static void cleanup_init_ggtt(struct i915_ggtt *ggtt)mutex_destroy(&ggtt->error_mutex); } +static void+ggtt_insert_scratch_pages(struct i915_ggtt *ggtt, u64 offset, u64 length)+{ + struct i915_address_space *vm = &ggtt->vm; + + if (GRAPHICS_VER(ggtt->vm.i915) < 8) + return vm->clear_range(vm, offset, length); + /* clear_range since gen8 is nop */It would also work to simply drop the <gen8 case and just do the loop below, right? If so would that be clearer?
Alternatively, if you want to keep the dual path here, perhaps it would be better to replace the gen check with "clear_range !=/== nop_clear_range". That way maybe more of the platform knowledge remains at a central location.
Another thing - are there any suspend-resume considerations (i915_ggtt_resume_vm)? Maybe this padding needs to be restored too.
Regards, Tvrtko
+ while (length > 0) {Maybe add a GEM_BUG_ON if length is not aligned? Granted it may be a huge stretch of imagination..Regards, Tvrtko+ vm->insert_page(vm, px_dma(vm->scratch[0]), offset, I915_CACHE_NONE, 0);+ offset += I915_GTT_PAGE_SIZE; + length -= I915_GTT_PAGE_SIZE; + } +} + static int init_ggtt(struct i915_ggtt *ggtt) { /* @@ -551,8 +566,12 @@ static int init_ggtt(struct i915_ggtt *ggtt) * paths, and we trust that 0 will remain reserved. However, * the only likely reason for failure to insert is a driver * bug, which we expect to cause other failures... + * + * Since CPU can perform speculative reads on error capture + * (write-combining allows it) add scratch page after error + * capture to avoid DMAR errors. */ - ggtt->error_capture.size = I915_GTT_PAGE_SIZE; + ggtt->error_capture.size = 2 * I915_GTT_PAGE_SIZE; ggtt->error_capture.color = I915_COLOR_UNEVICTABLE; if (drm_mm_reserve_node(&ggtt->vm.mm, &ggtt->error_capture)) drm_mm_insert_node_in_range(&ggtt->vm.mm, @@ -562,11 +581,15 @@ static int init_ggtt(struct i915_ggtt *ggtt) 0, ggtt->mappable_end, DRM_MM_INSERT_LOW); } - if (drm_mm_node_allocated(&ggtt->error_capture)) + if (drm_mm_node_allocated(&ggtt->error_capture)) { + u64 start = ggtt->error_capture.start; + u64 size = ggtt->error_capture.size; + + ggtt_insert_scratch_pages(ggtt, start, size); drm_dbg(&ggtt->vm.i915->drm, "Reserved GGTT:[%llx, %llx] for use by error capture\n", - ggtt->error_capture.start, - ggtt->error_capture.start + ggtt->error_capture.size); + start, start + size); + } /* * The upper portion of the GuC address space has a sizeable hole
